Skip to main content
Risk lists tool calls that endpoint detectors flagged, with an AI assessment of each call. Use the findings to decide whether you need a new policy. After an endpoint detector flags a call, a hosted AI assessment assigns a severity and category and writes a short explanation. Detection covers both steps and what data the assessment receives.

​
Read the headline

The dashboard shows “N risky calls of M flagged for review” at the top of Risk. The flagged count includes all calls that detectors flagged in the selected time range. The risky count includes calls the AI assessment rated as risky. The sentence under it ends with “A flag does not block: only a policy does, and policies blocked X of these calls.” The blocked count includes calls that a policy stopped. In Observe mode, a call runs even if the policy returns Deny. The dashboard shows Deny in the Decision column but excludes the call from the blocked count. The assessment rates what the call requested, regardless of the outcome. A blocked call can still be critical.

​
Cards

Kontext updates counts as assessments finish after the call. Not rated yet covers calls still waiting for an assessment and calls whose assessment failed.

​
Flagged tool calls

The table lists every flagged call in the time range, with the columns Severity, What happened, Agent, Who, Decision, and Time. What happened is the first sentence of the AI assessment. Decision shows the policy decision: Allow or Deny. Filter by Decision, Severity, Category, or Agent. The Decision filter offers Allowed and Blocked based on the endpoint’s outcome. Blocked shows only calls a policy stopped.

​
Open a call

Select a row to open the tool call drawer, the same one Explore uses. It has five tabs: Verdict, Risk, Input, Output, and Context. The Risk tab shows the AI assessment, marked advisory · never enforced. It includes the explanation, severity, categories, and evidence strength. It also lists the category’s ATT&CK, ATLAS, and OWASP entries. Detection details shows results from Kestrel, Merlin, and the Additional shell check.

​
Investigate a finding

  1. Read the request on the Verdict tab and the assessment on the Risk tab. Identify the resource, destination, or account the call would affect.
  2. Compare the Decision column with the result on the Verdict tab. A Deny in Observe mode means the call ran. An error in the result means the tool attempted the action and failed.
  3. The Context tab shows who the agent acted for, which endpoint it ran on, and its session. Open that session in Explore to review earlier tool output the agent read. Instructions hidden in a web page or file can cause a risky call.
  4. If the call affected a real system, check that system too.
For example, an agent runs git push --force origin main, and the assessment rates it high. The Block GitHub force pushes preset matched in Observe mode, so the Decision column shows Deny but the push ran. Check the repository, then move the policy to Enforce to stop the next attempt. See add a policy and roll out a policy.

​
Give feedback on Kestrel

When Kestrel scored the call, the Risk tab shows Kestrel feedback with two buttons: Fine and Should’ve blocked. Admins can save or change a label. Kontext stores it with the call without changing the decision or your policies. To get flagged findings in Slack, set up risk alerts.